Re: Bug: ip utility fails to show routes with large # of multipath next-hops

[David Ahern <dsahern@gmail.com>]

On 7/29/20 5:43 AM, Ido Schimmel wrote:
> On Tue, Jul 28, 2020 at 05:52:44PM -0700, Ashutosh Grewal wrote:
>> Hello David and all,
>>
>> I hope this is the correct way to report a bug.
> 
> Sure
> 
>>
>> I observed this problem with 256 v4 next-hops or 128 v6 next-hops (or
>> 128 or so # of v4 next-hops with labels).
>>
>> Here is an example -
>>
>> root@a6be8c892bb7:/# ip route show 2.2.2.2
>> Error: Buffer too small for object.
>> Dump terminated
>>
>> Kernel details (though I recall running into the same problem on 4.4*
>> kernel as well) -
>> root@ubuntu-vm:/# uname -a
>> Linux ch1 5.4.0-33-generic #37-Ubuntu SMP Thu May 21 12:53:59 UTC 2020
>> x86_64 x86_64 x86_64 GNU/Linux
>>
>> I think the problem may be to do with the size of the skbuf being
>> allocated as part of servicing the netlink request.
>>
>> static int netlink_dump(struct sock *sk)
>> {
>>   <snip>
>>
>>                 skb = alloc_skb(...)
> 
> Yes, I believe you are correct. You will get an skb of size 4K and it
> can't fit the entire RTA_MULTIPATH attribute with all the nested
> nexthops. Since it's a single attribute it cannot be split across
> multiple messages.

yep, well known problem.

> 
> Looking at the code, I think a similar problem was already encountered
> with IFLA_VFINFO_LIST. See commit c7ac8679bec9 ("rtnetlink: Compute and
> store minimum ifinfo dump size").
> 
> Maybe we can track the maximum number of IPv4/IPv6 nexthops during
> insertion and then consult it to adjust 'min_dump_alloc' for
> RTM_GETROUTE.

That seems better than the current design for GETLINK which walks all
devices to determine max dump size. Not sure how you will track that
efficiently though - add is easy, delete is not.

> 
> It's a bit complicated for IPv6 because you can append nexthops, but I
> believe anyone using so many nexthops is already using RTA_MULTIPATH to
> insert them, so we can simplify.

I hope so.

> 
> David, what do you think? You have a better / simpler idea? Maybe one
> day everyone will be using the new nexthop API and this won't be needed
> :)

exactly. You won't have this problem with separate nexthops since each
one is small (< 4k) and the group (multipath) is a set of ids, not the
full set of attributes.